Impact localization in composite structures with Deep Neural Networks

Abstract

This research looks at the application of Deep Neural Networks (DNNs) for low-energy impact localization in composite structures, a key aspect of structural health monitoring in the aerospace sector. The methodology used in this study involves the generation of a consistent impact dataset using an autonomous impact machine, followed by meticulous data processing. The training of the DNN models was focused on minimizing the Euclidean distance between the predicted and actual impact positions employing custom loss functions. This study yielded several significant findings. First, it confirmed the feasibility of using DNNs for effective impact localization in complex composite structures, although with varying degrees of accuracy across different impact locations but with an average error of the same order as the labeling error. Second, it was observed that the performance of the models was considerably influenced by structural features, such as the presence of stringers and the placement of sensors. The architecture demonstrated consistent performance across multiple trained models, indicating their robustness and potential for generalization. The implications of these findings for structural health monitoring are substantial, suggesting that DNNs can be a valuable tool for early damage detection in composite structures.